Human-Like Storyteller: A Hierarchical Network with Gated Memory for Visual Storytelling

摘要

Different from the visual captioning that describes an image concretely, the visual storytelling aims at generating an imaginative paragraph with a deep understanding of the given image stream. It is more challenging for the requirements of inferring contextual relationships among images. Intuitively, humans tend to tell the story around a central idea that is constantly expressed with the continuation of the storytelling. Therefore, we propose the Human-Like StoryTeller (HLST), a hierarchical neural network with a gated memory module, which imitates the storytelling process of human beings. First, we utilize the hierarchical decoder to integrate the context information effectively. Second, we introduce the memory module as the story's central idea to enhance the coherence of generated stories. And the multi-head attention mechanism with a self adjust query is employed to initialize the memory module, which distils the salient information of the visual semantic features. Finally, we equip the memory module with a gated mechanism to guide the story generation dynamically. During the generation process, the expressed information contained in memory is erased with the control of the read and write gate. The experimental results indicate that our approach significantly outperforms all state-of-the-art (SOTA) methods.